I have a new toy to play with (a Quadro RTX) and I want to use CUDA.jl with the GPU.
I get a quite simple yet annoying issue associated to the function cached_memory.
The very simple MWE
julia> CUDA.versioninfo()
CUDA toolkit 11.4.1, artifact installation
CUDA driver 11.2.0
NVIDIA driver 460.91.3
Libraries:
- CUBLAS: 11.5.4
- CURAND: 10.2.5
- CUFFT: 10.5.1
- CUSOLVER: 11.2.0
- CUSPARSE: 11.6.0
- CUPTI: 14.0.0
- NVML: 11.0.0+460.91.3
- CUDNN: 8.20.2 (for CUDA 11.4.0)
- CUTENSOR: 1.3.0 (for CUDA 11.2.0)
Toolchain:
- Julia: 1.6.2
- LLVM: 11.0.1
- PTX ISA support: 3.2, 4.0, 4.1, 4.2, 4.3, 5.0, 6.0, 6.1, 6.3, 6.4, 6.5, 7.0
- Device capability support: sm_35, sm_37, sm_50, sm_52, sm_53, sm_60, sm_61, sm_62, sm_70, sm_72, sm_75, sm_80
1 device:
0: Quadro RTX 5000 (sm_75, 14.922 GiB / 15.738 GiB available)
julia> CUDA.cached_memory()
ERROR: CUDA error: invalid argument (code 1, ERROR_INVALID_VALUE)
Stacktrace:
[1] throw_api_error(res::CUDA.cudaError_enum)
@ CUDA ~/.julia/packages/CUDA/VGl9W/lib/cudadrv/error.jl:105
[2] macro expansion
@ ~/.julia/packages/CUDA/VGl9W/lib/cudadrv/error.jl:115 [inlined]
[3] cuMemPoolGetAttribute(pool::CuMemoryPool, attr::CUDA.CUmemPool_attribute_enum, value::Base.RefValue{UInt64})
@ CUDA ~/.julia/packages/CUDA/VGl9W/lib/utils/call.jl:26
[4] attribute(X::Type, pool::CuMemoryPool, attr::CUDA.CUmemPool_attribute_enum)
@ CUDA ~/.julia/packages/CUDA/VGl9W/lib/cudadrv/pool.jl:77
[5] cached_memory()
@ CUDA ~/.julia/packages/CUDA/VGl9W/src/pool.jl:532
[6] top-level scope
@ REPL[8]:100:
The main issue is that the function cached_memory() is used several times (for instance for allocation) at several places (like in cudnnFindConvolutionAlgorithmWorkspaceSize for conv in GPU)
Any hints to address that ?
